Add BNNM intrinsic.

Change-Id: I05fc9835786455e28d69f2b4ffe70c3292a01404
diff --git a/rs/java/android/renderscript/RenderScript.java b/rs/java/android/renderscript/RenderScript.java
index 126b8c7..71d0a5b 100644
--- a/rs/java/android/renderscript/RenderScript.java
+++ b/rs/java/android/renderscript/RenderScript.java
@@ -949,6 +949,17 @@
         rsnScriptIntrinsicBLAS_Z(mContext, id, func, TransA, TransB, Side, Uplo, Diag, M, N, K, alphaX, alphaY, A, B, betaX, betaY, C, incX, incY, KL, KU);
     }
 
+    native void rsnScriptIntrinsicBLAS_BNNM(long con, long id, int M, int N, int K,
+                                             long A, int a_offset, long B, int b_offset, long C, int c_offset,
+                                             int c_mult_int);
+    synchronized void nScriptIntrinsicBLAS_BNNM(long id, int M, int N, int K,
+                                             long A, int a_offset, long B, int b_offset, long C, int c_offset,
+                                             int c_mult_int) {
+        validate();
+        rsnScriptIntrinsicBLAS_BNNM(mContext, id, M, N, K, A, a_offset, B, b_offset, C, c_offset, c_mult_int);
+    }
+
+
 
     long     mDev;
     long     mContext;
diff --git a/rs/java/android/renderscript/ScriptIntrinsicBLAS.java b/rs/java/android/renderscript/ScriptIntrinsicBLAS.java
index 90d2300..16b7033 100644
--- a/rs/java/android/renderscript/ScriptIntrinsicBLAS.java
+++ b/rs/java/android/renderscript/ScriptIntrinsicBLAS.java
@@ -176,6 +176,9 @@
     private static final int RsBlas_zherk = 141;
     private static final int RsBlas_zher2k = 142;
 
+    // BLAS extensions start here
+    private static final int RsBlas_bnnm = 1000;
+
     /**
      */
     public static ScriptIntrinsicBLAS create(RenderScript rs) {
@@ -1485,5 +1488,23 @@
     }
 
 
+    /**
+     *
+     * 8-bit GEMM-like operation for neural networks
+     *
+     * @hide
+     **/
+    public void BNNM(Allocation A, int a_offset, Allocation B, int b_offset, Allocation C, int c_offset, int c_mult) {
+        validateL3(Element.U8(mRS), NO_TRANSPOSE, TRANSPOSE, 0, A, B, C);
+
+        int M = -1, N = -1, K = -1;
+        M = A.getType().getY();
+        N = B.getType().getY();
+        K = A.getType().getX();
+
+
+        mRS.nScriptIntrinsicBLAS_BNNM(getID(mRS), M, N, K, A.getID(mRS), a_offset, B.getID(mRS), b_offset, C.getID(mRS), c_offset, c_mult);
+
+    }
 
 }
diff --git a/rs/jni/android_renderscript_RenderScript.cpp b/rs/jni/android_renderscript_RenderScript.cpp
index 676d94f..078f65b 100644
--- a/rs/jni/android_renderscript_RenderScript.cpp
+++ b/rs/jni/android_renderscript_RenderScript.cpp
@@ -582,6 +582,32 @@
 
 
 static void
+nScriptIntrinsicBLAS_BNNM(JNIEnv *_env, jobject _this, jlong con, jlong id, jint M, jint N, jint K,
+                                             jlong A, jint a_offset, jlong B, jint b_offset, jlong C, jint c_offset,
+                                             jint c_mult_int) {
+    RsBlasCall call;
+    memset(&call, 0, sizeof(call));
+    call.func = RsBlas_bnnm;
+    call.M = M;
+    call.N = N;
+    call.K = K;
+    call.a_offset = a_offset;
+    call.b_offset = b_offset;
+    call.c_offset = c_offset;
+    call.c_mult_int = c_mult_int;
+
+    RsAllocation in_allocs[3];
+    in_allocs[0] = (RsAllocation)A;
+    in_allocs[1] = (RsAllocation)B;
+    in_allocs[2] = (RsAllocation)C;
+
+    rsScriptForEachMulti((RsContext)con, (RsScript)id, 0,
+                         in_allocs, sizeof(in_allocs), nullptr,
+                         &call, sizeof(call), nullptr, 0);
+}
+
+
+static void
 nAssignName(JNIEnv *_env, jobject _this, jlong con, jlong obj, jbyteArray str)
 {
     if (kLogApi) {
@@ -2425,6 +2451,8 @@
 {"rsnScriptIntrinsicBLAS_Complex",   "(JJIIIIIIIIIFFJJFFJIIII)V",             (void*)nScriptIntrinsicBLAS_Complex },
 {"rsnScriptIntrinsicBLAS_Z",         "(JJIIIIIIIIIDDJJDDJIIII)V",             (void*)nScriptIntrinsicBLAS_Z },
 
+{"rsnScriptIntrinsicBLAS_BNNM",      "(JJIIIJIJIJII)V",                       (void*)nScriptIntrinsicBLAS_BNNM },
+
 {"rsnProgramStoreCreate",            "(JZZZZZZIII)J",                         (void*)nProgramStoreCreate },
 
 {"rsnProgramBindConstants",          "(JJIJ)V",                               (void*)nProgramBindConstants },